Document 10779532

advertisement
15th Annual Conference of Applied Mathematics, Univ. of Central Oklahoma,
Electronic Journal of Differential Equations, Conference 02, 1999, pp. 1–9.
ISSN: 1072-6691. URL: http://ejde.math.swt.edu or http://ejde.math.unt.edu
ftp ejde.math.swt.edu (login: ftp)
Fifth-order Runge-Kutta with higher order
derivative approximations ∗
David Goeken & Olin Johnson
Abstract
0
Given y = f (y), standard Runge-Kutta methods perform multiple
evaluations of f (y) in each integration sub-interval as required for a given
accuracy. Evaluations of y 00 = fy f or higher derivatives are not considered
due to the assumption that the calculations involved in these functions
exceed those of f . However, y 00 can be approximated to sufficient accuracy from past and current evaluations of f to achieve a higher order of
accuracy than is available through current functional evaluations alone.
In July of 1998 at the ANODE (Auckland Numerical Ordinary Differential Equations) Workshop, we introduced a new class of Runge-Kutta
methods based on this observation (Goeken 1999). We presented a thirdorder method which requires only two evaluations of f and a fourth-order
method which requires three. This paper reviews these two methods and
gives the general solution to the equations generated by the fifth-order
methods of this new class. Interestingly, these fifth-order methods require
only four functional evaluations per step whereas standard Runge-Kutta
methods require six.
1
Third-order method
We consider initial value problems expressed in autonomous form. Starting with
the non-autonomous form, we assume that f (x, y) is a continuous function with
domain D in Rn+1 where x ∈ R, y ∈ Rn and (x, y) ∈ D We assume that
kf (x, y1 ) − f (x, y2 )k2 ≤ Lky1 − y2 k2
for all (x, y1 ), (x, y2 ) ∈ D; thus the problem
y 0 = f (x, y)
y(x0 ) = y0 with (x0 , y0 ) ∈ D
has a unique solution.
∗ 1991 Mathematics Subject Classifications: 65L06.
Key words and phrases: multistep Runge-Kutta, third-order method,
fourth-order method, fifth-order method, higher order derivatives.
c
1999
Southwest Texas State University and University of North Texas.
Published November 23, 1999.
1
2
Fifth-order Runge-Kutta
In autonomous form, y and f have n + 1 components with yn+1 = x and
fn+1 (y) = 1. The initial value problem is then written
y 0 = f (y)
y(x0 ) = y0 where (y0 )n+1 = x0 .
Most efforts to increase the order of the Runge-Kutta methods have been accomplished by increasing the number of Taylor’s series terms used and thus the
number of functional evaluations required (Butcher 1987) (Gear 1971). The use
of higher order derivative terms has been proposed for stiff problems (Rosenbrock 1963) (Enright 1974). Our method adds higher order derivative terms to
the Runge-Kutta ki terms (i > 1) to achieve a higher order of accuracy. For
example, our new third-order method, GJ3, for autonomous systems, lets
yn+1 = yn + b1 k1 + b2 k2
and k1 = hf (yn ). However, we introduce additional terms by assigning
k2 = hf (yn + a21 k1 + ha22 fy (yn )k1 )
Using Taylor’s series expansion techniques, the above is uniquely satisfied to
O(h3 ) as follows
2
k1
=
k2
=
yn+1
=
hf (yn )
2
2
hf (yn + k1 + hfy (yn )k1 )
3
9
1
3
yn + k1 + k2
4
4
Utilizing fy
The previous section developed a two-stage, third-order method; however, it
introduced a term with fy . The result is the addition of a higher derivative term
to the standard Runge-Kutta method. The following describes three methods
to utilize fy .
Method 1: If one knows or can generate fy , and if the evaluation of fy is
cheaper than the evaluation of f , then savings can be realized. For example,
with a linear system of equations, y 0 = Ay, fy is known and constant.
Method 2: Since y 00 = f 0 = fy f for autonomous equations, and since k1 = hf ,
k2 can be replaced with
k2
=
=
2
hf (yn + k1 +
3
2
hf (yn + k1 +
3
2
hfy k1 )
9
2
hfy hf )
9
David Goeken & Olin Johnson
2
= hf (yn + k1 +
3
2
= hf (yn + k1 +
3
3
2 2
h fy f )
9
2 2 0
h f)
9
or
2
2
k2 = hf (yn + k1 + h2 y 00 ) .
3
9
Again, savings can be realized if one can formulate y 00 (or f 0 ) and if it is cheaper
to evaluate than f .
Method 3: Building onto Method 2, one can approximate y 00 (or f 0 ) by using
the current and previous evaluations of f . For our third-order method, this
approximation must be of O(h). Since an O(h) approximation of f 0 is given by
f 0 = (fn − fn−1 )/h, one can approximate k2 as follows
k2
=
=
=
2
hf (yn + k1 +
3
2
hf (yn + k1 +
3
2
hf (yn + k1 +
3
2 2 0
h f)
9
2 2
h (fn − fn−1 )/h)
9
2
h(fn − fn−1 ))
9
Since fn is calculated in the current step in the evaluation of k1 , one only has
to store the previous value, fn−1 . In effect, the use of previous values for the
approximation has created a multistep Runge-Kutta method.
3
Fourth-order method
Similarly, our fourth-order method, GJ4, for autonomous systems, lets
yn+1 = yn + b1 k1 + b2 k2 + b3 k3
and
k1
k2
= hf (yn )
= hf (yn + a21 k1 + ha22 fy (yn )k1 )
k3
= hf (yn + a31 k1 + a32 k2 + ha33 fy (yn )k1 + ha34 fy (yn )k2 )
The Taylor’s series expansion of these higher order methods is tedious and error
prone. We used modern symbolic math packages to expand and then to solve the
resulting systems of nonlinear equations that were generated. In this work, we
used the symbolic math packages Reduce (Reduce 1999), PARI/GP (PARI/GP
1999), and Octave (Eaton 1997).
PARI/GP was used to generate the Taylor’s series expansion of the above,
resulting in the following system of equations
b1 + b2 + b3 = 1
4
Fifth-order Runge-Kutta
b2 a21 + b3 [a31 + a32 ] = 1/2
b2 a221 + b3 [a31 + a32 ]2 = 1/3
b2 a321 + b3 [a31 + a32 ]3 = 1/4
b2 a22 + b3 [a21 a32 + a33 + a34 ] = 1/6
b3 [a21 a34 + a22 a32 ] = 1/24
1
b2 a21 a22 + b3 [a21 a32 ( a21 + a31 + a32 ) + (a31 + a32 )(a33 + a34 )] = 1/6
2
However, in order to utilize Methods 2 and 3 of Section 2, we must restrict the
solution with a34 = 0. The general solution to the above system of equations
(with a34 = 0) has been found with Reduce and example solutions are shown in
Table 1.
Table 1: Example of fourth-order autonomous solutions
b1
1/6
1/6
1/6
1/10
1/10
4
b2
1/6
2/3
2/3
1/2
1/2
b3
2/3
1/6
1/6
2/5
2/5
a21
1
1/2
1/2
1/3
1/3
a22
1/2
1/8
−1/8
1/18
−1/6
a31
3/8
-1
3
−25/24
35/24
a32
1/8
2
-2
15/8
−5/8
a33
0
−1/2
5/2
−5/18
5/6
Fifth-order method
In July of 1998, the authors presented (Goeken 1999) this numerical integration
technique at a meeting attended by Dr. John Butcher. Using his tree-based
approach (Butcher 1987), Dr. Butcher suggested a fifth-order method. Since
the meeting, his technique has been verified using Taylor’s series expansion
techniques to determine the general solution for our fifth-order methods.
Our fifth-order method, GJ5, for autonomous systems, lets
yn+1 = yn + b1 k1 + b2 k2 + b3 k3 + b4 k4
and
k1
=
hf (yn )
k2
k3
=
=
hf (yn + a21 k1 + ha22 fy (yn )k1 )
hf (yn + a31 k1 + a32 k2 + ha33 fy (yn )k1 )
k4
=
hf (yn + a41 k1 + a42 k2 + a43 k3 + ha44 fy (yn )k1 )
PARI/GP was used to generate the Taylor’s series expansion of the above,
resulting in the following system of equations
b1 + b2 + b3 + b4 = 1
David Goeken & Olin Johnson
5
b2 a21 + b3 [a31 + a32 ] + b4 [a41 + a42 + a43 ] = 1/2
b2 a221 + b3 [a31 + a32 ]2 + b4 [a41 + a42 + a43 ]2 = 1/3
b2 a22 + b3 [a21 a32 + a33 ] + b4 [a21 a42 + a43 (a31 + a32 ) + a44 ] = 1/6
b2 a321 + b3 [a31 + a32 ]3 + b4 [a41 + a42 + a43 ]3 = 1/4
1
1
b2 a21 a22 + b3 [ a221 a32 + (a31 + a32 )(a21 a32 + a33 )] + b4 [a221 a42
2
2
+a43 (a31 + a32 )2 + 2(a41 + a42 + a43 )(a21 a42 + (a31 + a32 )a43 + a44 )] = 1/6
b3 a22 a32 + b4 [a21 a32 a43 + a22 a42 + a33 a43 ] = 1/24
b2 a421 + b3 [a31 + a32 ]4 + b4 [a41 + a42 + a43 ]4 = 1/5
3b2 a221 a22 + b3 [a321 a32 + 3(a31 + a32 )2 (a21 a32 + a33 )] + b4 [a321 a42
+(a31 + a32 )3 a43 + 3(a41 + a42 + a43 )2 (a21 a42 + (a31 + a32 )a43 + a44 )] = 7/20
b3 a221 a32 (a31 + a32 ) + b4 [(a41 + a42 + a43 )(a221 a42 + (a31 + a32 )2 a43 )] = 1/15
1
1
1
b2 a222 + b3 [a21 a32 ( a21 a32 + a22 + a33 ) + a22 a32 (a31 + a32 ) + a233 ]
2
2
2
1 2
+b4 [ a21 (a32 a43 + a242 ) + (a31 + a32 )(a21 (a32 a43 + a42 a43 ) + a43 (a33 + a44 )
2
1
+ (a31 + a32 )a243 ) + a21 a42 (a22 + a44 ) + (a21 a32 a43 + a22 a42
2
1
+a33 a43 )(a41 + a42 + a43 ) + a244 ] = 11/120
2
b4 a22 a32 a43 = 1/120
The solution presented by Dr. Butcher and verified using the above system of
equations is
k1
=
k2
=
k3
=
k4
=
yn+1
=
hf (yn )
1
1
hf (yn + k1 + hfy k1 )
3
18
152
252
44
k1 +
k2 −
hfy k1 )
hf (yn −
125
125
125
19
72
25
5
hf (yn + k1 − k2 + k3 + hfy k1 )
2
7
14
2
5
27
125
1
k3 + k4
yn + k1 + k2 +
48
56
336
24
Additional solutions to the above nonlinear system of equations have been found
using Octave. Three additional solutions are shown in Table 2.
5
Numerical results
To demonstrate that the new methods are of the order claimed, several equations have been solved using the new third-, fourth-, and fifth-order method
6
Fifth-order Runge-Kutta
Table 2: Example of fifth-order autonomous solutions
b1
b2
b3
b4
a21
a22
a31
a32
a33
a41
a42
a43
a44
1/24
125/336
27/56
5/48
1/5
1/50
-52/27
70/27
-8/27
43/5
-64/7
54/35
13/10
5/54
250/567
32/81
1/14
3/10
9/200
-9/8
15/8
-9/32
17/3
-490/81
112/81
23/18
1/14
32/81
250/567
5/54
1/4
1/32
-329/250
252/125
-259/1000
209/35
-32/5
10/7
11/10
Table 3: Test problems
Function
y 0 = −y
y 0 = y4 −
y2
80
Solution
y = e−t
20
y=
−t
1+19e
4
y(0)
1
1
on scalar autonomous equations, systems of autonomous equations, and scalar
non-autonomous equations.
Our previous paper (Goeken 1999) demonstrated the third- and fourth-order
methods utilizing fy and the approximation to f 0 . Here we will concentrate
on our new fifth-order method. For scalar autonomous examples, we use the
equations shown in Table 3, with initial condition y(0) = 1. These equations
were solved using a standard fifth-order Runge-Kutta method along with our
fifth-order method, GJ5, using fy directly. Relative error was plotted against
the step size and is shown in Figures 1 and 2. Results are comparable to
standard fifth-order Runge-Kutta solution, thus demonstrating our claim. The
new method requires four functional evaluations of f and one of fy per step or
four functional evaluations of f and three historical values of f ; whereas, the
standard fifth-order Runge-Kutta method requires six functional evaluations of
f.
6
Conclusions
New third-, fourth-, and fifth-order numerical integration techniques inspired
by the Runge-Kutta method have been presented. The new methods exploit
David Goeken & Olin Johnson
7
Figure 1 - Equation y’=-y
1e+00
Runge-Kutta 5
GJ 5
Relative Error at t=20
1e-02
1e-04
1e-06
1e-08
1e-10
0.032
0.064
0.128
0.256
Step Size
0.512
1.024
the use of higher order derivatives, specificly fy . In particular, a technique
utilizing an approximation to y 00 has been presented resulting in a multistep
Runge-Kutta method. Table 4 compares the computational effort required for
standard Runge-Kutta methods with our methods. Table 4 shows the cases
where the proposed methods are more efficient than the standard Runge-Kutta
methods. Specifically, the proposed methods are more efficient for cases where
• fy or y 00 is cheaper to evaluate than f ,
• the use of historical values of f is cheaper then evaluating f , and
• for the fifth-order case, the number of total functional evaluations can be
reduced from 6 to 4 when using an approximation of f 0 .
Table 4: Number of evaluations comparison
Order
3
4
5
Standard
Runge-Kutta
Num. f
Evals.
3
4
6
Proposed Method
Exact fy
Num. f Num. fy
Evals
Evals.
2
1
3
1
4
1
Proposed Method
Approximating f 0
Num. f Num. fn−i
Evals.
Values
2
1
3
2
4
3
8
Fifth-order Runge-Kutta
Figure 2 - Equation y’=y*(1/4-y/80)
1e-06
Runge-Kutta 5
GJ 5
Relative Error at t=20
1e-08
1e-10
1e-12
1e-14
1e-16
0.032
0.064
0.128
0.256
Step Size
0.512
1.024
References
[1] Butcher, J. C. (1987) The Numerical Analysis of Ordinary Differential Equations Runge-Kutta and General Linear Methods, John Wiley & Sons Ltd.,
New York
[2] Eaton, J. W. (1997) GNU Octave A high-level interactive language for numerical computations,
http://www.we.fh-osnabrueck.de/labsim/octave/manual/octave.html
[3] Enright, W. H. (1974) Second Derivative Multistep Methods for Stiff Ordinary Differential Equations, SIAM J. Numer. Anal., 11, 321-331
[4] Gear, C. W. (1971) Numerical Initial Value Problems in Ordinary Differential Equations, Prentice-Hall, Englewood Cliffs, New Jersey
[5] Goeken, D. and Johnson, O. (1999) Runge-Kutta with Higher Order Derivative Approximations, submitted to Applied Numerical Mathematics
[6] PARI/GP (1999)
http://hasse.mathematik.tu-muenchen.de/ntsw/pari/Welcome
[7] Reduce (1999)
http://www.sub.uni-goettingen.de/ssgfi/math/infodata/000838.html
[8] Rosenbrock, H. H. (1963) Some General Implicit Processes for the Numerical
Solution of Differential Equations, Comp. J., 5, 329-330
David Goeken & Olin Johnson
David Goeken
Department of Computer Science
The University of Houston
Houston, TX 77204-3475, USA
e-mail: dgoeken@cs.uh.edu
Now with the LinCom Corporation, Houston, Texas
Olin Johnson
Department of Computer Science
The University of Houston
Houston, TX 77204-3475, USA
e-mail: johnson@cs.uh.edu
9
Download